Picture the planning meeting. Someone from sales wants an agent who chases every lead. Service wants one that resolves cases end-to-end. Finance wants one that handles billing disputes. IT notes that no one has reviewed the knowledge base in 2 years. Every idea is reasonable, the budget covers one or two of them, and the loudest voice in the room is about to decide which.
That is how most companies pick their first AI agent: by enthusiasm, seniority, or whichever demo looked best. It is also why so many first deployments stall. The idea was good. The choice of where to start was not.
This guide offers a different way to decide. Instead of scoring everything on one scale and picking the winner, you run candidates through three gates, rank the ones that survive, and then decide how much freedom the agent gets on day one. We call it Gate, Rank, Ladder.
Why the Obvious Ranking Is Usually the Wrong One
Ask a leadership team to rank agent ideas, and they will almost always sort by potential payoff. Autonomous billing disputes beat help-desk answers. Contract negotiation beats meeting summaries. The bigger the prize, the higher it climbs.
The trouble is that the payoff only matters if the agent works. High-value tasks tend to share three awkward traits: they touch money or legal commitments, they depend on data scattered across several systems, and nobody has measured how long they take today. Start there, and you are attempting the hardest version of the problem with the least preparation.
A better first deployment is one you can do well, learn from quickly, and defend to a skeptical finance team. Payoff still counts. It just shouldn’t be the first question.
Step One: Three Gates
Before any idea gets a score, it has to pass three yes-or-no gates. Failing a gate doesn’t kill the idea. It moves it to a later wave once the missing piece is fixed.
Gate 1: Can we afford to be wrong?
Every agent will be wrong sometimes. The question is what happens next.
- Reversible: a wrong answer can be spotted and corrected before it matters, such as a misrouted case or a poor draft email.
- Costly but contained: a mistake causes rework or an apology, but not lasting harm.
- Irreversible: money has moved, a contract term has changed, or a regulated decision has been made.
Only the first two pass. Agents that make irreversible decisions don’t go first.
Gate 2: Is the data ready?
An agent is only as reliable as what it reads. Check three things for the task in question:
- The information it needs exists and is reasonably accurate
- It can be reached through the systems you already run
- Permissions are set up so the agent sees only what the person they serve is allowed to see
That last point is easy to overlook. If your org uses controlled access by hiding fields on page layouts rather than proper field-level security, an agent reading data directly will not respect those hiding. It is worth checking before launch, not after. (Our piece on the agentic UI and security gaps explains why.)
Gate 3: Can we measure it?
If you cannot say how long the task takes today, how often it goes wrong, or what it costs, you cannot prove an agent helped. Pick candidates where a baseline exists or can be captured in a couple of weeks.
An idea that fails the measurement gate isn’t bad. It just needs a baseline first, and capturing one is a useful small project on its own.
Step Two: Rank the Survivors
For ideas that pass all three gates, a simple formula separates the strong from the merely interesting:
Priority = (Value × Reach) ÷ Effort
Rate each on a scale of 1 to 5:
- Value: how much the business gains if it works (time saved, revenue protected, experience improved)
- Reach: how many people or interactions it affects, and how often
- Effort: how much work to build, connect, test, and roll out (a higher number means more effort)
Dividing by effort is the important part. It rewards quick, contained projects and keeps expensive ones from winning on ambition alone.
Â
Five Candidates Through the Process
Everything below is hypothetical, picked to show how the method behaves. Your scores should come from your own people and data.
Â
|
CANDIDATE |
GATE 1: SAFE TO BE WRONG? |
GATE 2: DATA READY? |
GATE 3: MEASURABLE? |
VALUE |
REACH |
EFFORT |
PRIORITY |
|
C. Internal IT and HR help desk |
Pass |
Pass |
Pass |
3 |
4 |
1 |
12.0 |
|
A. Case triage and routing |
Pass |
Pass |
Pass |
4 |
5 |
2 |
10.0 |
|
B. Lead follow-up drafting |
Pass |
Pass |
Pass |
4 |
4 |
3 |
5.3 |
|
D. Autonomous billing disputes |
Fail |
Partial |
Pass |
5 |
3 |
4 |
held (3.75 if it passed) |
|
E. Autonomous pricing negotiation |
Fail |
Partial |
Fail |
5 |
2 |
5 |
held |
The arithmetic for the three that pass:
- C: (3 × 4) ÷ 1 = 12.0
- A: (4 × 5) ÷ 2 = 10.0
- B: (4 × 4) ÷ 3 = 5.33
What the table says
The help desk and case triage go first. The help desk has the lowest effort and a safe failure mode, so it is the best place to learn how agents behave in your environment. Case triage has more reach and a visible customer benefit, and its failure mode is also mild: a misrouted case gets rerouted.
Lead follow-up goes second. The payoff is real, but it takes more effort, usually because of contact and activity data that needs tidying.
Billing disputes and pricing negotiation wait. They carry the largest potential payoff on the board, and they still do not belong in the first wave. Both involve irreversible outcomes. Each can come back later in a different form: an agent that prepares a dispute summary for a person to approve, or one that gathers pricing context for a negotiator, passes Gate 1, where full autonomy does not.
That last move is worth remembering. A failed gate often points to a smaller version of the same idea that does pass.
Step Three: The Autonomy Ladder
Choosing where to start is half the decision. The other half is how much freedom the agent gets there. Think of it as a ladder with five rungs, and treat each climb as something earned with evidence.
Â
|
RUNG |
WHAT THE AGENT DOES |
WHO DECIDES |
TYPICAL USE |
|
1. Suggest |
Offers an answer or next step |
A person, entirely |
Knowledge lookup, case summaries |
|
2. Draft |
Prepares the work product |
A person edits and sends |
Emails, replies, notes |
|
3. Act with approval |
Proposes an action and waits |
A person approves of each one |
Record updates, routing, and refunds under review |
|
4. Act within limits |
Acts alone inside set boundaries |
A person samples and audits |
Routine cases under a value or category cap |
|
5. Act broadly |
Handles a wide range of tasks independently |
Monitoring and exceptions |
Rarely appropriate early |
Â
Most first deployments should start on rung 1 or 2 and move up only when the data supports it. A reasonable test for climbing: enough real cases have been handled, the error rate is known and acceptable, and escalation to a person works smoothly.
Pairing the ladder with the ranking gives each candidate a starting point:
- Help desk (C): begin at rung 2, move to rung 4 for simple, low-stakes requests
- Case triage (A): begin at rung 3, since routing mistakes are cheap to reverse
- Lead follow-up (B): stay at rung 2, since outbound messages carry brand risk
- Billing disputes (D): only at rung 1 or 2 for now, as a preparation aid
- Pricing (E): rung 1 only, as research support
A Sample 90 Days
Timing varies a great deal with company size and complexity, so treat this as a shape, not a schedule.
Days 1 to 30: Choose and prepare. Run the gates with the people who do the work, not only leadership. Capture baselines for the top two candidates. Audit permissions and the content the agent will rely on. Name an owner for each use case.
Days 31 to 60: Pilot on a short ladder. Launch the first agent at rung 1 or 2 with a small group. Review every output at the start. Log errors by type. Collect user feedback weekly.
Days 61 to 90: Decide with evidence. Compare results to the baseline. Decide whether to climb a rung, fix the data, widen the audience, or stop. Start the second candidate only after the first has a clear verdict.
Warning Signs a Pilot Is Going Sideways
You do not need a formal review to spot trouble. Watch for these:
- Users stop checking the agent’s work because it has been right a few times, then miss an error
- People route around the agent and keep doing the task the old way
- Nobody can say who owns the agent’s performance
- The same type of mistake keeps recurring, and nothing changes
- Success is described in feelings (“people like it”) instead of numbers
- The scope keeps widening before the original task is complete
- Knowledge articles the agent cites are out of date
Any one of these is a prompt to pause and fix, not to push on.
Notes for the United States, the UAE, and the United Kingdom
The method stays the same across regions. What shifts is the weight of the data and the compliance gate.
United States. Requirements depend heavily on sector and state. Healthcare, financial services, and public-sector work carry their own rules, and customers often ask for evidence of security controls during procurement. Have your compliance team confirm what applies before a customer-facing pilot.
United Arab Emirates. Data protection and, in some sectors, data residency expectations can shape where an agent may run and what it may read. Confirm requirements with your legal advisers and check your platform’s current regional hosting options. Bilingual (Arabic and English) support can also affect the effort score for customer-facing agents.
United Kingdom. UK GDPR places clear expectations on how personal data is used, how automated decisions are explained, and when a person must be involved. Include data protection input at Gate 2, and be especially careful about agents that influence decisions affecting individuals.
These are pointers for conversation with your own advisers. They are not legal advice, and the rules change, so confirm the current position before you rely on it.
What to do on the 1st Day
- List every agent idea your teams have raised, including the ones you doubt.
- Run each one through the three gates with the people who do the work.
- Rank the survivors with the priority formula.
- Pick the top one or two, and decide which ladder rung each starts on.
- Capture the baseline before anything launches.
- Write down what result would make you climb a rung, and what would make you stop.
The first agent you deploy matters less for its own payoff than for what it teaches your organization about data, oversight, and trust. Choose the one that teaches those lessons cheaply, and the bigger ideas get much easier to approve.
If you would like a second pair of eyes on your shortlist, our Salesforce team can run the gates with you in a short working session.
